Deep learning systems explore the interior of materials from the outside
(Nanowerk News) Maybe you can’t tell a book by its cover, but according to researchers at MIT, you can now do the equivalent for all kinds of materials, from airplane parts to medical implants. Their new approach allows engineers to know what’s going on inside simply by observing the surface properties of the material.
The team used a type of machine learning known as deep learning to compare large sets of simulated data about the material’s external force fields and their corresponding internal structures, and used that to come up with a system that can make reliable predictions of the surface’s interior. data.
The results are published in a journal Advanced Materials (“Fill in the Blanks: A Transferrable Deep Learning Approach to Recovering Lost Physical Field Information”), in a paper by doctoral student Zhenze Yang and professor of civil and environmental engineering Markus Buehler.
“This is a very common problem in engineering,” explains Buehler. “If you have a material — maybe a car door or an airplane — and you want to know what’s inside that material, you can measure the tension in the surface by taking a picture and calculating how much deformation you’re experiencing. own. But you can’t really see into the material. The only way you can do that is to cut it open and then look inside and see if there’s any damage there.
It is also possible to use X-rays and other techniques, but these tend to be expensive and require large equipment, he said. “So, what we’re basically doing is asking the question: Can we develop an AI algorithm that can see what’s going on on the surface, which we can easily see by either using a microscope or taking a photo, or maybe just measuring something on the surface of a material, and then try to figure out what’s really going on inside?” The inside information may include any damage, cracks or stresses to the material, or details of its internal microstructure.
The same types of questions can also be applied to biological networks, he added. “Is there any disease there, or some sort of growth or change in the tissue?” The goal is to develop a system that can answer these kinds of questions in a completely non-invasive way.
Achieving that goal involves dealing with complexity including the fact that “many problems like that have multiple solutions,” says Buehler. For example, many different internal configurations may exhibit the same surface properties. To address that ambiguity, “we’ve created a method that can give us all the possibilities, all the options, basically, that might generate this particular (surface) scenario.”
The technique they developed involved training an AI model using large amounts of data about surface measurements and the interior properties associated with them. This includes not only materials that are uniform but also materials that differ in combination. “Some of the newer planes are made of composites, so they have a deliberate design to have different phases,” said Buehler. “And of course, in biology as well, any kind of biological material is going to be made of lots of components and they have very different properties, like in bones, where you have very soft proteins, and then you have very rigid mineral substances.”
The technique works even for materials whose complexity isn’t fully understood, he says. “With complex biological networks, we don’t understand exactly how it behaves, but we can measure its behavior. We don’t have the theory for it, but if we have enough data collected we can train the model.”
That said the method they developed is widely applicable. “This is not only limited to solid mechanics problems, but can also be applied to various engineering disciplines, such as fluid dynamics and other types.” Buehler added that it could be applied to determine a wide range of properties, not just stresses and strains, but fluid fields or magnetic fields, for example the magnetic field inside a fusion reactor. It is “quite universal, not only for different materials, but also for different disciplines.”
Yang said that he initially started thinking about this approach when he was studying data on materials in which parts of the image he was using were blurred, and he wondered how it would be possible to “fill in the blanks” from the missing data in the blurred areas. “How can we recover this lost information?” he wondered. Reading further, he discovered that this was an example of a widespread problem, known as the reverse problem, of trying to recover lost information.
Developing a method involves an iterative process, building the model making initial predictions, comparing them with actual data on the material in question, then refining the model further to fit that information. The resulting model was tested against cases where the material was understood enough to be able to calculate its true internal properties, and the new method’s predictions matched those calculated properties.
The training data includes surface images, but also various other types of measurements of surface properties, including pressure, and electric and magnetic fields. In many cases, researchers use simulated data based on an understanding of the underlying structure of certain materials. And even when a new material has many unknown characteristics, the method can still produce good enough estimates to provide engineers with general directions on how to make further measurements.
As an example of how this methodology can be applied, Buehler points out that currently, aircraft are often inspected by testing a few representative areas with expensive methods such as X-rays because it is impractical to test the entire aircraft. “It’s a different approach, where you have a much cheaper way to collect data and make predictions,” said Buehler. “From there you can then make a decision about where you want to look, and possibly use more expensive equipment to test it.”
First of all, he expects this method, which is freely available for anyone to use via the GitHub website, to have wide application in laboratory environments, for example in test materials used for soft robotics applications.
For such materials, he said, “We can measure something on the surface, but we don’t know what is often going on inside the material, because it’s made of hydrogels or proteins or biomaterials for actuators, and there’s no theory. therefore. So that’s an area where researchers can use our techniques to make predictions about what’s going on inside, and maybe design better grippers or composites.”
